3

I am coming over from Keras to PyTorch and am having a hard time creating graph visualizations of models. In Keras, you can simply call plot_model(), and that function will use graphviz to produce a nice graph.

However, in PyTorch it is not so easy. I've seen there are a lot of tools available, such as TorchViz, but they all require that you pass in some input into the model. For example:

  1. This StackOverflow question (How do I visualize a net in Pytorch?) suggests using the torchviz package like so:
x = torch.zeros(1, 3, 224, 224, dtype=torch.float, requires_grad=False)
out = resnet(x)
make_dot(out)
  1. The HiddenLayer package (example) also requires dummy input:
model = torchvision.models.vgg16()
g = hl.build_graph(model, torch.zeros([1, 3, 224, 224]))

So my question is: Suppose I download a pre-trained model that has little documentation. How can I determine what is the needed dummy input dimension in order to run these visualization functions?

For example, when I print() an AlexNet model, what part of the result below tells me that the input should be [1, 3, 224, 224]?

model = torchvision.models.alexnet(pretrained=True)
print(model)

# AlexNet(
#   (features): Sequential(
#     (0): Conv2d(3, 64, kernel_size=(11, 11), stride=(4, 4), padding=(2, 2))
#     (1): ReLU(inplace=True)
#     (2): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
#     (3): Conv2d(64, 192, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
#     (4): ReLU(inplace=True)
#     (5): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
#     (6): Conv2d(192, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
#     (7): ReLU(inplace=True)
#     (8): Conv2d(384, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
#     (9): ReLU(inplace=True)
#     (10): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
#     (11): ReLU(inplace=True)
#     (12): MaxPool2d(kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False)
#   )
#   (avgpool): AdaptiveAvgPool2d(output_size=(6, 6))
#   (classifier): Sequential(
#     (0): Dropout(p=0.5, inplace=False)
#     (1): Linear(in_features=9216, out_features=4096, bias=True)
#     (2): ReLU(inplace=True)
#     (3): Dropout(p=0.5, inplace=False)
#     (4): Linear(in_features=4096, out_features=4096, bias=True)
#     (5): ReLU(inplace=True)
#     (6): Linear(in_features=4096, out_features=1000, bias=True)
#   )
# )

The first layer says: Conv2d(3, 64, kernel_size=(11, 11), stride=(4, 4), padding=(2, 2)). How would I know from that fact that I need to create [1, 3, 224, 224] input?

stackoverflowuser2010
  • 38,621
  • 48
  • 169
  • 217
  • As far as I know, if you don't know the models' input shape, you need to deduce that from the own model. Looking only at the first layer isn't possible to get the shape. But, if you know the remaining layers you can do that. Of course, it may be tiresome. Does Keras compute it automatically? For an unknown model? – André Pacheco Mar 11 '20 at 11:34
  • @AndréPacheco: In Keras, the model implementer has to define the input dimensions exactly, and it will be displayed. https://keras.io/models/model/ – stackoverflowuser2010 Mar 11 '20 at 20:07
  • So, PyTorch doesn't demand that. You need to do what I said in the previous reply – André Pacheco Mar 12 '20 at 14:02

0 Answers0